Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Basis (linear algebra) wikipedia , lookup
Hilbert space wikipedia , lookup
Fundamental group wikipedia , lookup
Covering space wikipedia , lookup
Bra–ket notation wikipedia , lookup
Group action wikipedia , lookup
Invariant convex cone wikipedia , lookup
AN APPLICATION OF A FUNCTIONAL INEQUALITY TO QUASI-INVARIANCE IN INFINITE DIMENSIONS MARIA GORDINA† Abstract. One way to interpret smoothness of a measure in infinite dimensions is quasi-invariance of the measure under a class of transformations. Usually such settings lack a reference measure such as the Lebesgue or Haar measure, and therefore we can not use smoothness of a density with respect to such a measure. We describe how a functional inequality can be used to prove quasi-invariance results in several settings. In particular, this gives a different proof of the classical Cameron-Martin (Girsanov) theorem for an abstract Wiener space. In addition, we revisit several more geometric examples, even though the main abstract result concerns quasi-invariance of a measure under a group action on a measure space. Contents 1. Introduction Acknowledgement 2. Notation 3. Finite-dimensional approximations and quasi-invariance 4. A functional inequality 5. Examples 5.1. Abstract Wiener space 5.2. Wang’s Harnack inequality 5.3. Infinite-dimensional Heisenberg-like groups: References 1 2 2 3 5 7 7 10 11 13 1. Introduction Our goal in this paper is to describe how a functional inequality can be used to prove quasi-invariance of certain measures in infinite dimensions. Even though the original argument was used in a geometric setting, we take a slightly different approach in this paper. Namely, we formulate a method that can be used to prove quasi-invariance of a measure under a group action. Such methods are useful in infinite dimensions when usually there is no a natural reference measure such as the Lebesgue measure. At the same time quasi-invariance of measures is a useful tool in proving regularity results when it is reformulated as an integration by parts formula. We do not discuss significance of such results, and 1991 Mathematics Subject Classification. Primary 58G32 58J35; Secondary 22E65 22E30 22E45 58J65 60B15 60H05. Key words and phrases. Quasi-invariance; group action; functional inequalities. † Research was supported in part by NSF Grant DMS-1007496. 1 2 GORDINA moreover we do not refer to the extensive literature on the subject, as it is beyond the scope of our paper. We start by describing an abstract setting of how finite-dimensional approximations can be used to prove such a quasi-invariance. In [9] this method was applied to projective and inductive limits of finite-dimensional Lie groups acting on themselves by left or right multiplication. In that setting a functional inequality (integrated Harnack inequality) on the finite-dimensional approximations leads to a quasi-invariance theorem on the infinite-dimensional group space. Similar methods were used in the elliptic setting on infinite-dimensional Heisenberg-like groups in [8], and on semi-infinite Lie groups in [14]. Note that the assumptions we make below in Section 3 have been verified in these settings, including the sub-elliptic case for infinite-dimensional Heisenberg group in [1]. Even though the integrated Harnack inequality we use in these situations have a distinctly geometric flavor, we show in this paper that it does not have to be. The paper is organized as follows. The general setting is described in Section 2 and 3, where Theorem 3.2 is the main result. One of the ingredients for this result is quasi-invariance for finite-dimensional approximations which is described in Section 3. We review the connection between an integrated Harnack inequality and Wang’s Harnack inequality in Section 4. Finally, Section 5 gives several examples of how one can use Theorem 3.2. We describe in detail the case of an abstract Wiener space, where the group in question is identified with the Cameron-Martin subspace acting by translation on the Wiener space. In addition we discuss elliptic (Riemannian) and sub-elliptic (sub-Riemannian) infinite-dimensional groups which are examples of a subgroup acting on the group by multiplication. Acknowledgement. The author is grateful to Sasha Teplyaev and Tom Laetsch for useful discussions and helpful comments. 2. Notation Suppose G is a topological group with the identity e, X is a topological space, (X, B, µ) is a measure space, where B is the Borel σ-algebra, and µ is a probability measure. We assume that G is endowed with the structure of a Hilbert Lie group (e.g. [7]), and further that its Lie algebra g := Lie (G) = Te G is equipped with a Hilbertian inner product, ⟨·, ·⟩. The corresponding distance on G is denoted by d (·, ·). In addition, we assume that G is separable, and therefore we can use what is known about Borel actions of Polish groups [2,3]. Once we have an inner product on the Lie algebra g, we can define the length of a path in G as follows. Suppose k ∈ C 1 ([0, 1], G), k (0) = e, then ∫ (2.1) · lG (k (·)) := 0 |Lk(t)−1 ∗ k̇ (t) |dt, where Lg is the left translation by g ∈ G. We assume that G acts measurably on X, that is, there is a (Borel) measurable map Φ : G × X −→ X such that Φ (e, x) = x, for all x ∈ X, Φ (g1 , Φ (g2 , x)) = Φ (g1 g2 , x) , for all x ∈ X, g1 , g2 ∈ G. AN APPLICATION OF A FUNCTIONAL INEQUALITY 3 We often will use Φg := Φ (g, ·) for g ∈ G. Definition 2.1. Suppose Φ is a measurable group action of G on X. (1) In this case we denote by (Φg )∗ µ the pushforward measure defined by ( ( )) (Φg )∗ µ (A) := µ Φ g −1 , A , for all A ∈ B, g ∈ G; (2) the measure µ is invariant under the action Φ if (Φg )∗ µ = µ for all g ∈ G; (3) the measure µ is quasi-invariant with respect to the action Φ if (Φg )∗ µ and µ are mutually absolutely continuous for all g ∈ G. Notation 2.2. For a topological group G acting measurably on the measure space (X, B, µ) in such a way that µ is quasi-invariant under the action by G, the RadonNikodym derivative of (Φg )∗ µ with respect to µ is denoted by Jg (x) := (Φg )∗ µ (dx) for all g ∈ G, x ∈ X. µ (dx) For a thorough discussion of the Radon-Nikodym derivative in this setting we refer to [5, Appendix D] 3. Finite-dimensional approximations and quasi-invariance We start by describing approximations to both the group G and the measure space (X, B, µ). At the end we also need to impose certain conditions to have consistency of the group action defined on these approximations. As X is a topological space, we denote by Cb (X) the space of continuous bounded real-valued functions. Assumption 1 (Lie group assumptions). Suppose {Gn }n∈N is a collection of finitedimensional unimodular Lie subgroups of G such that Gn ⊂ Gm for all n < m. We assume that there exists a smooth section {sn : G −→ Gn }n∈N , that ∪ is, sn ◦ in = idGn , where in : Gn −→ G is the smooth injection. We suppose that n∈N Gn is a dense subgroup of G. In addition, we assume that the length of a path in G can be approximated by the lengths in Gn , namely, if k ∈ C 1 ([0, 1], G), k (0) = e, then (3.1) lG (k (·)) = lim lGn (sn (k (·))) . n→∞ Note that sn does not have to be a group homomorphism. Assumption 2 (Measure space assumptions). We assume that X is a separable topological space with a sequence of topological spaces Xn ⊂ X which come with corresponding continuous maps πn : X −→ Xn satisfying the following properties. For any f ∈ Cb (X) ∫ ∫ f ◦ jn dµn , f dµ = lim (3.2) X n→∞ Xn where jn : Xn −→ X is the continuous injection map, and µn is the pushforward measure (πn )∗ µ. Our last assumption concerns the group action for these approximations. 4 GORDINA Assumption 3 (Group action assumptions). The approximations to group G and the measure space (X, µ) are consistent with the group action in the following way Φ (Gn × Xn ) ⊂ Xn for each n ∈ N, Φg : X → X is a continuous map for each g ∈ G. We denote by Φn the restriction of Φ to Gn × Xn . Observe that Φn = Φ ◦ (in , jn ) which together with Assumption 3, it is clear that Φn is a measurable group action of Gn on (Xn , Bn , µn ). Suppose now that µn is quasi-invariant under the group action Φn , and let Jgn ( ) be the Radon-Nikodym derivative Φng ∗ µn with respect to µn . We assume that there is a positive constant C = C (p) such that for any p ∈ [1, ∞) and g ∈ Gn ( ) ∥Jgn ∥Lp (Xn ,µn ) 6 exp C (p) d2Gn (e, g) . (3.3) Note that the constant C (p) does not depend on n. Remark 3.1. The fact that this estimate is Gaussian (with the square of the distance) does not seem to be essential. But as we do not have examples with a different exponent, we leave (3.3) as is. Moreover, we could consider a more general function on the right hand side than an exponential of the distance squared. Theorem 3.2 (Quasi-invariance of µ). Suppose we have a group G and a measure space (X, B, µ) satisfying Assumptions 1, 2 and 3, and the uniform estimate (3.3) on the Radon-Nikodym derivatives holds. Then for all g ∈ G the measure µ is quasi-invariant under the action Φg . Moreover, for all p ∈ (1, ∞), d (Φg )∗ µ ( ) (3.4) 6 exp C (p) d2G (e, g) . dµ p L (X,µ) Proof. Using (3.3) we see that for any bounded continuous function f ∈ Cb (X), n ∈ N, and g ∈ G ∫ ∫ n |(f ◦ jn )(Φsn (g) (x))|dµn (x) = Jsnn (g) (x)|(f ◦ jn )(x)|dµn (x) Xn Xn ( ) 6 ∥f ◦ jn ∥Lp′ (Xn ,µn ) exp C (p) d2Gn (e, sn (g)) , where p′ is the conjugate exponent to p. Note that by Assumption 3 and definitions of jn and Φn for all (g, x) ∈ Gn × Xn ( ) jn Φng (x) = Φng (x) = Φg (x) = Φg (jn (x)) and therefore ( ) f ◦ jn (Φng (x)) = f Φng (x) = f ◦ Φg (jn (x)) , (g, x) ∈ Gn × Xn . Thus ∫ ( ) | (f ◦ jn ) Φng (x) |dµn (x) = Xn ∫ | (f ◦ Φg ) (jn (x)) |dµn (x). Xn Allowing n → ∞ in the last identity and using (3.2) and the fact that f ◦Φg ∈ Cb (X) yields ∫ ( ) |f (Φg (x)) |dµ(x) 6 ∥f ∥Lp′ (X,µ) exp C (p) d2Gn (e, g) , for all g ∈ Gn . (3.5) X AN APPLICATION OF A FUNCTIONAL INEQUALITY 5 Thus, we have proved that (3.5) holds for f ∈ Cb (X) and g ∈ Gn . Now we would like to prove (3.5) for the distance dG instead of dGn with g still in Gn . Take any path k ∈ C 1 ([0, 1], G) such that k (0) = e and k (1) = g, and observe that then sn ◦ k ∈ C 1 ([0, 1], Gn ) and therefore (3.5) holds with d2Gn (e, g) replaced by lGn (sn ◦ k) (1). Now we can use (3.2) in Assumption 1 and optimizing over all such paths k to see that (3.6)∫ ∪ ) ( Gn . |f (Φg (x)) |dµ(x) 6 ∥f ∥Lp′ (X,µ) exp C (p) d2G (e, g) , for all g ∈ X n∈N By Assumption 1 this union is dense in G, therefore dominated convergence along with the continuity of dG (e, g) in g implies that (3.6) holds for all g ∈ G. Since the ′ bounded continuous functions are dense in Lp (X, µ) (see for example [11, Theorem A.1, p. 309]), the inequality in (3.6) implies that the linear functional φg : Cb (X) → R defined by ∫ |f (Φg (x)) |dµ(x) φg (f ) := X ′ has a unique extension to an element, still denoted by φg , of Lp (X, µ)∗ which satisfies the bound ) ( |φg (f )| 6 ∥f ∥Lp′ (X,µ) exp C (p) d2G (e, g) ′ ′ for all f ∈ Lp (X, µ). Since Lp (X, µ)∗ ∼ = Lp (W, µ), there exists a function Jg ∈ p L (X, µ) such that ∫ (3.7) φg (f ) = f (x)Jg (x)dµ(x), X ′ for all f ∈ Lp (X, µ), and ( ) ∥Jg ∥Lp (X,µ) 6 exp C (p) d2G (e, g) . Now restricting (3.7) to f ∈ Cb (X), we may rewrite this equation as ∫ ∫ (3.8) f (Φg (x)) dµ(x) = f (x)Jg (x)dµ(x). X W Then a monotone class argument (again use [11, Theorem A.1]) shows that (3.8) is valid for all bounded measurable functions f on W . Thus, d (Φg )∗ µ/dµ exists and is given by Jg , which is in Lp for all p ∈ (1, ∞) and satisfies the bound (3.4). 4. A functional inequality In this section we would like to revisit an observation made in [9]. Namely, [9, Lemma D.1] connects Wang’s Harnack inequality with an estimate similar to 3.3. It is easy to transfer this argument from the setting of Riemannian manifolds to a more general situation. We start with an integral operator on L2 (X, ν), where (X, ν) is a σ-finite measure space. Namely, let ∫ T f (x) := p (x, y) f (y) dν (y) , f ∈ L2 (X, ν) , X 6 GORDINA where the integral kernel p (x, y) is assumed to satisfy the following properties. positive p (x, y) > 0 for all x, y ∈ X, ∫ p (x, y) dν (y) = 1 for all x ∈ X, conservative X symmetric p (x, y) = p (y, x) for all x, y ∈ X, continuous p (·, ·) : X × X −→ R is continuous. Some of these assumptions might not be needed for the proof of Proposition 4.1, but we make them to simplify the exposition. Note that in our applications this integral kernel is the heat kernel for a strongly continuous, symmetric, Markovian semigroup in L2 (X, ν), therefore the corresponding heat kernel is positive, symmetric with the total mass not exceeding 1, in addition to having the semigroup property or being the approximate identity in L2 (X, ν). In our examples this heat semigroup is also conservative, therefore the heat kernel is conservative (stochastically complete), and thus p (x, y) dν (y) is a probability measure. The following proposition is a generalization of [9, Lemma D.1], and it simply ′ ∗ reflects the fact that (Lp ) and Lp are isometrically isomorphic Banach spaces for 1 < p < ∞ and p′ = p/ (p − 1), the conjugate exponent to p. Proposition 4.1. Let x, y ∈ X, p ∈ (1, ∞) and C ∈ (0, ∞] which might depend on x and y. Then p [(T f ) (x)] 6 C p (T f p ) (y) for all f > 0 (4.1) if and only if (∫ [ (4.2) X p (x, z) p (y, z) )1/p′ ]p ′ p (y, z) dν (z) 6 C. Proof. Since p (·, ·) is positive, we can write ∫ p (x, z) (T f ) (x) = f (z) p (y, z) dν (z) . X p (y, z) We denote dµy (·) := p (y, ·) dν (·) and gx,y (·) := p(x,·) p(y,·) , then ∫ (4.3) (T f ) (x) = f (z) gx,y (z) dµy (z) . X ∗ ′ Since gx,y > 0 and L (µ) is isomorphic to Lp (µ), the pairing in (4.3) implies that ∫ f (z) gx,y (z) dµy (z) (T f ) (x) = sup . ∥gx,y ∥Lp′ (µ) = sup X 1/p ∥f ∥ p f >0 [(T f p ) (y)] f >0 L (µy ) p The last equation may be written more explicitly as (∫ [ )1/p′ ]p ′ p (x, z) (T f ) (x) p (y, z) dν (z) = sup , 1/p p (y, z) f >0 X [(T f p ) (y)] and from this equation the result follows. AN APPLICATION OF A FUNCTIONAL INEQUALITY 7 Remark 4.2. In the case when T is a Markov semigroup Pt and p (·, ·) = pt (·, ·) is the corresponding integral kernel, Proposition 4.1 shows that a Wang’s Harnack inequality is equivalent to an integrated Harnack inequality. Subsection 5.2 gives more details on this equivalence for Riemannian manifolds, see Corollary 5.10. Remark 4.3. The connection between Proposition 4.1 and (3.3) can be seen if we choose x and y in (4.2) as the endpoints of the group action as follows. Let x, y ∈ X and g ∈ G be such that Φe (x) = x and Φg (x) = y, then to apply Proposition 4.1 we can take the constant in (4.2) to be equal to ( ) exp C (p − 1) d2Gn (e, g) . Here the measure on X is dµx (z) = p (x, z) dν (z). 5. Examples 5.1. Abstract Wiener space. Standard references on basic facts on the Gaussian measures include [4, 12]. Let (H, W, µ) be an abstract Wiener space, that is, H is a real separable Hilbert space densely continuously embedded into a real separable Banach space W , and µ is the Gaussian measure defined by the characteristic functional ( ) ∫ |φ|2 ∗ eiφ(x) dµ (x) = exp − H 2 W for any φ ∈ W ∗ ⊂ H ∗ . We will identify W ∗ with a dense subspace of H such that for any h ∈ W ∗ the linear functional ⟨·, h⟩ extends continuously from H to W . We will usually write ⟨φ, w⟩ := φ (w) for φ ∈ W ∗ , w ∈ W . More details can be found in [4]. It is known that µ is a Borel measure, that is, it is defined on the Borel σ-algebra B (W ) generated by the open subsets of W . We would like to apply the material from Sections 4 3 with (X, µ) = (W, µ) and the group G = EW being the group of (measurable) rotations and translations by the elements from the Cameron-Martin subspace H. We can view this group as an infinite-dimensional analogue of the Euclidean group. Notation 5.1. We call an orthogonal transformation of H which is a topological homeomorphism of W ∗ a rotation of W ∗ . The space of all such rotations is denoted by O (W ). For any R ∈ O (W ∗ ) its adjoint, R∗ , is defined by ⟨φ, R∗ w⟩ := ⟨R−1 φ, w⟩, w ∈ W, φ ∈ W ∗ . Proposition 5.2. For any R ∈ O (W ) the map R∗ is a B (W )-measurable map from W to W and µ ◦ (R∗ ) −1 = µ. ∗ Proof. The measurability of R follows from the fact that R is continuous on H. For any φ ∈ W ∗ ∫ iφ(x) ( ∗ −1 ) ∫ i⟨φ,x⟩ ( ∗ −1 ) ∫ x = e dµ (R ) x = e dµ (R ) ei⟨φ,R W W W ( ) ( ) ∫ |R−1 φ|2H ∗ |φ|2H ∗ exp − = exp − = eiφ(x) dµ (x) 2 2 W ∗ x⟩ dµ (x) = 8 GORDINA since R is an isometry. Corollary 5.3. Any R ∈ O (W ) extends to a unitary map on L2 (W, µ). Definition 5.4. The Euclidean group EW is a group generated by measurable rotations R ∈ O (W ) and translation Th : W → W , Th (w) := w + h. To describe finite-dimensional approximations as in Section 3 we need to give more details on the identification of W ∗ with a dense subspace of H. Let i : H → W be the inclusion map, and i∗ : W ∗ → H ∗ be its transpose, i.e. i∗ ℓ := ℓ ◦ i for all ℓ ∈ W ∗ . Also let H∗ := {h ∈ H : ⟨·, h⟩H ∈ Ran(i∗ ) ⊂ H ∗ } or in other words, h ∈ H is in H∗ iff ⟨·, h⟩H ∈ H ∗ extends to a continuous linear functional on W . We will continue to denote the continuous extension of ⟨·, h⟩H to W by ⟨·, h⟩H . Because H is a dense subspace of W , i∗ is injective and because i is injective, i∗ has a dense range. Since h 7→ ⟨·, h⟩H as a map from H to H ∗ is a conjugate linear isometric isomorphism, it follows from the above comments that for any h ∈ H we have h 7→ ⟨·, h⟩H ∈ W ∗ is a conjugate linear isomorphism too, and that H∗ is a dense subspace of H. Now suppose that P : H → H is a finite rank orthogonal projection such that n P H ⊂ H∗ . Let {ej }j=1 be an orthonormal basis for P H and ℓj = ⟨·, ej ⟩H ∈ W ∗ . Then we may extend P to a (unique) continuous operator from W → H (still denoted by P ) by letting (5.1) Pn w := n ∑ ⟨w, ej ⟩H ej = j=1 n ∑ ℓj (w) ej for all w ∈ W. j=1 As we pointed put in [8, Equation 3.43] there exists C < ∞ such that ∥P w∥H 6 C ∥w∥W for all w ∈ W. (5.2) Notation 5.5. Let Proj (W ) denote the collection of finite rank projections on W such that P W ⊂ H∗ and P |H : H → H is an orthogonal projection, i.e. P has the form given in Equation (5.1). ∞ Also let {ej }j=1 ⊂ H∗ be an orthonormal basis for H. For n ∈ N, define Pn ∈ Proj (W ) as in Notation 5.5, i.e. (5.3) Pn (w) = n ∑ ⟨w, ej ⟩H ej = j=1 n ∑ ℓj (w) ej for all w ∈ W. j=1 Then we see that Pn |H ↑ IdH . Proposition 5.6. The Gaussian measure µ is quasi-invariant under the translations from H and invariant under orthogonal transformations of H. Proof. The second part of the statement is the content of Proposition 5.2. We now prove quasi-invariance of µ under translation by elements in H. Let {Pn }n∈N be a collection of operators defined by (5.3)for an orthonormal basis {ej }∞ j=1 of n ∼ H such that {ej }∞ ⊆ H . Then H := P (H) R , and the pushforward = ∗ n n j=1 measure (Pn )∗ µ is simply the standard Gaussian measure pn (x) dx on Hn . So if we identify the group of translation G with H and sn := Pn |H , then the group AN APPLICATION OF A FUNCTIONAL INEQUALITY 9 action is given by Φh (w) := w + h, w ∈ W, h ∈ H. Note that Assumptions 1, 2 and 3 are satisfied, where jn = Pn : W −→ Hn etc. In particular, if we denote hn := Pn (h) ∈ Rn , h ∈ H, then for any measurable function f : W −→ R we see that f ◦ Pn (w + h) = f ◦ Pn ◦ Φhn (w) = f ◦ Φhn ◦ Pn (w) . Therefore ∫ ∫ f ◦ Pn (w + h) dµ (w) = f (x + Pn h) pn (x) dx = W Hn ∫ ∫ f (x) pn (x − hn ) dx = Hn f (x) Hn pn (x − hn ) pn (x) dx pn (x) ∫ = f (x) Jhn (x) pn (x) dx. Hn Using an explicit form of the Radon-Nikodym derivative Jhn (x), we see that for ′ any f ∈ Lp (W, µ) ∫ W pn (x − hn ) |f ◦ Pn (w + h) |dµ (w) 6 ∥f ∥Lp′ (pn (x)dx) pn (x) p L (pn (x)dx) ( 6 ∥f ◦ Pn ∥Lp′ (pn (x)dx) exp (p − 1) ∥hn ∥2H 2 ) . Thus (3.3) is satisfied, and therefore Theorem 3.2 is applicable, which proves the quasi-invariance with the Radon-Nikodym derivative satisfying ( (5.4) ∥Jh ∥Lp (W,µ) 6 exp ) (p − 1) ∥h∥2H . 2 Remark 5.7. The statement of Proposition 5.6 of course follows from the CameronMartin theorem which states that µ is quasi-invariant under translations by elements in H with the Radon-Nikodym derivative given by ( ) |h|2 d µ ◦ Th−1 d (Th )∗ µ d (µ ◦ T−h ) (w) = (w) = (w) = e−⟨h,w⟩− 2 , w ∈ W, h ∈ H. dµ dµ dµ Thus (5.4) is sharp. Remark 5.8. Following [10] we see that quasi-invariance of the Gaussian measure µ induces the Gaussian regular representation of the Euclidean group EW on L2 (W, µ) by 10 GORDINA ( )1/2 ( ) d (µ ◦ (Th R∗ )) −1 (UR,h f ) (w) := (w) f (Th R∗ ) (w) = dµ ( )1/2 ( ) d (µ ◦ Th ) −1 (w) f (R∗ ) (w − h) = dµ ( ) |h|2 −1 e⟨h,w⟩− 2 f (R∗ ) (w − h) , w ∈ W which is well-defined by Corollary 5.3. It is clear that this is a unitary representation. 5.2. Wang’s Harnack inequality. This follows [9, Appendix D]. The following theorem appears in [15,16] with k = −K, V ≡ 0. We will use the following notation { (5.5) c (t) := t et −1 1 t ̸= 0, t = 0. Theorem 5.9 (Wang’s Harnack inequality). Suppose that M is a complete connected Riemannian manifold such that Ric > kI for some k ∈ R. Then for all p > 1, f > 0, t > 0, and x, y ∈ M we have ( ′ ) pk 2 p (5.6) (Pt f ) (y) 6 (Pt f p ) (z) exp kt d (y, z) . e −1 Corollary 5.10. Let (M, g) be a complete Riemannian manifold such that Ric > kI for some k ∈ R. Then for every y, z ∈ M and p ∈ [1, ∞) (∫ [ ( ]p )1/p ) pt (y, x) c (kt) (p − 1) 2 (5.7) 6 exp pt (z, x) dV (x) d (y, z) 2t M pt (z, x) where c (·) is defined by (5.5), pt (x, y) is the heat kernel on M and d (y, z) is the Riemannian distance from x to y for x, y ∈ M . Proof. From Lemma 4.1 and Theorem 5.9 with ) ( ) ( ′ k 1 k p 2 2 d (y, z) = exp d (y, z) , C = exp p ekt − 1 p − 1 ekt − 1 it follows that it follows that )1/p′ (∫ [ ]p′ ( ) 1 k pt (x, z) 2 pt (y, z) dV (z) 6 exp d (y, z) . p − 1 ekt − 1 M pt (y, z) Using p−1 = (p′ − 1) −1 and then interchanging the roles of p and p′ gives (5.7). The reason we call 4.2 an integrated Harnack inequality on a d-dimensional manifold M is as follows. Recall the classical Li–Yau Harnack inequality ( [13] and [6, Theorem 5.3.5]) which states that if α > 1, s > 0, and Ric > −K for some K > 0, then ( )dα/2 ) ( 2 pt (y, x) t+s dαKs αd (y, z) (5.8) 6 + , exp pt+s (z, x) t 2s 8 (α − 1) for all x, y, z ∈ M and t > 0. However, when s = 0, (5.8) gives no information on pt (y, x) /pt (z, x) when y ̸= z. This inequality is based on the Laplacian ∆/2 AN APPLICATION OF A FUNCTIONAL INEQUALITY 11 rather than ∆, t and s should be replaced by t/2 and s/2 when applying the results in [6, 13]. 5.3. Infinite-dimensional Heisenberg-like groups: Riemannian and subRiemannian cases. These examples represent infinite-dimensional versions of the group action of a Lie group on itself by left or right multiplication. The difference is in geometry of the space on which the group acts on: Riemannian and subRiemannian. In both cases we proved (3.3), where the constant C depends on the geometry, and the distance used is Riemannian or Carnot-Carathéodory. Let (W, H, µ) be an abstract Wiener space and let C be a finite-dimensional inner product space. Define g := W × C to be an infinite-dimensional Heisenberglike Lie algebra, which is constructed as an infinite-dimensional step 2 nilpotent Lie algebra with continuous Lie bracket. Namely, let ω : W × W → C be a continuous skew-symmetric bilinear form on W . We will also assume that ω is surjective. Let g denote W × C when thought of as a Lie algebra with the Lie bracket given by (5.9) [(X1 , V1 ), (X2 , V2 )] := (0, ω(X1 , X2 )). Let G denote W × C when thought of as a group with multiplication given by 1 g1 g2 := g1 + g2 + [g1 , g2 ], 2 where g1 and g2 are viewed as elements of g. For gi = (wi , ci ), this may be written equivalently as ( ) 1 (5.10) (w1 , c1 ) · (w2 , c2 ) = w1 + w2 , c1 + c2 + ω(w1 , w2 ) . 2 Then G is a Lie group with Lie algebra g, and G contains the subgroup GCM = H × C which has Lie algebra gCM . In terms of Section 2 the Cameron-Martin (Hilbertian) subgroup GCM is the group that is acting on the Heisenberg group G by left or right multiplication. Using Notation 5.5 we can define finite-dimensional approximations to G by using P ∈ Proj(W ). We assume in addition that P W is sufficiently large to satisfy Hörmander’s condition (that is, {ω(A, B) : A, B ∈ P W } = C). For each P ∈ Proj(W ), we define GP := P W × C ⊂ H∗ × C and a corresponding projection πP : G → GP πP (w, x) := (P w, x). We will also let gP = Lie(GP ) = P W × C. For each P ∈ Proj(W ), GP is a finite-dimensional connected unimodular Lie group. Notation 5.11. (Riemannian and horizontal distances on GCM ) (1) For x = (A, a) ∈ GCM , let |x|2gCM := ∥A∥2H + ∥a∥2C . The length of a C 1 -path σ : [0, 1] → GCM is defined as ∫ 1 ℓ(σ) := |Lσ−1 (s)∗ σ̇(s)|gCM ds. 0 By 1 CCM we denote the set of paths σ : [0, 1] → GCM . 12 GORDINA (2) A C 1 -path σ : [0, 1] → GCM is horizontal if Lσ(t)−1 ∗ σ̇(t) ∈ H × {0} for 1,h a.e. t. Let CCM denote the set of horizontal paths σ : [0, 1] → GCM . (3) The Riemannain distance between x, y ∈ GCM is defined by 1 d(x, y) := inf{ℓ(σ) : σ ∈ CCM such that σ(0) = x and σ(1) = y}. (4) The horizontal distance between x, y ∈ GCM is defined by 1,h dh (x, y) := inf{ℓ(σ) : σ ∈ CCM such that σ(0) = x and σ(1) = y}. The Riemannian and horizontal distances are defined analogously on GP and will be denoted by dP and dhP correspondingly. In particular, for a sequence {Pn }∞ n=1 ⊂ Proj(W ), we will let Gn := GPn , dn := dPn , and dhn := dhPn . Now we are ready to define the corresponding heat kernel measures on G. We start by considering two Brownian motions on g bt := (B (t) , B0 (t)) , t > 0, bht := (B (t) , 0 (t)) , t > 0, with variance determined by [ ] E ⟨(B (s) , B0 (s)) , (A, a)⟩gCM ⟨(B (t) , B0 (t)) , (C, c)⟩gCM = Re ⟨(A, a) , (C, c)⟩gCM min (s, t) for all s, t ∈ [0, ∞), A, C ∈ H∗ and a, c ∈ C. A (Riemannian) Brownian motion on G is the continuous G–valued process defined by ( ) ∫ 1 t (5.11) g (t) = B (t) , B0 (t) + ω (B (τ ) , dB (τ )) . 2 0 Further, for t > 0, let µt = Law (g (t)) be a probability measure on G. We refer to µt as the time t heat kernel measure on G. Similarly a horizontal Brownian motion on G is the continuous G–valued process defined by ( ) ∫ 1 t (5.12) g h (t) = B (t) , ω (B (τ ) , dB (τ )) . 2 0 ( ) Then for t > 0, let µht = Law g h (t) be a probability measure on G. We refer to µt as the time t horizontal heat kernel measure on G. As the proof of [8, Theorem 8.1] explains, in this case Assumptions 1, 2 and 3 are satisfied, and moreover, (3.3) is satisfied as follows. Namely, [8, Corollary 7.3] says that the Ricci curvature is bounded from below by k (ω) uniformly for all Gn , so (3.3) holds as follows. ( (5.13) ∥Jkn ∥Lp (µn ) 6 exp ) c (k (ω) t) (p − 1) 2 dn (e, k) , k ∈ Gn , 2t where c (·) is defined by (5.5). In the sub-Riemannian case we have AN APPLICATION OF A FUNCTIONAL INEQUALITY (5.14) h,n Jk (( Lp (µn h) 6 exp 8∥ω∥22,n 1+ ρ2,n ) 13 ( )2 ) (1 + p) dhn (e, k) , 4t where the geometric constants are defined as in [1, p. 25]. References [1] Fabrice Baudoin, Maria Gordina, and Tai Melcher. Quasi-invariance for heat kernel measures on sub-Riemannian infinite-dimensional Heisenberg groups. Trans. Amer. Math. Soc., 365(8):4313–4350, 2013. [2] Howard Becker. Polish group actions: dichotomies and generalized elementary embeddings. J. Amer. Math. Soc., 11(2):397–449, 1998. [3] Howard Becker and Alexander S. Kechris. Borel actions of Polish groups. Bull. Amer. Math. Soc. (N.S.), 28(2):334–341, 1993. [4] Vladimir I. Bogachev. Gaussian measures, volume 62 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1998. [5] Theo Bühler. On the algebraic foundation of bounded cohomology. PhD thesis, ETH, 2008. [6] E. B. Davies. Heat kernels and spectral theory, volume 92 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1989. [7] Pierre de la Harpe. Classical Banach-Lie algebras and Banach-Lie groups of operators in Hilbert space. Lecture Notes in Mathematics, Vol. 285. Springer-Verlag, Berlin-New York, 1972. [8] Bruce K. Driver and Maria Gordina. Heat kernel analysis on infinite-dimensional Heisenberg groups. J. Funct. Anal., 255(9):2395–2461, 2008. [9] Bruce K. Driver and Maria Gordina. Integrated Harnack inequalities on Lie groups. J. Differential Geom., 83(3):501–550, 2009. [10] Bruce K. Driver and Brian C. Hall. The energy representation has no non-zero fixed vectors. In Stochastic processes, physics and geometry: new interplays, II (Leipzig, 1999), volume 29 of CMS Conf. Proc., pages 143–155. Amer. Math. Soc., Providence, RI, 2000. [11] Svante Janson. Gaussian Hilbert spaces, volume 129 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1997. [12] Hui Hsiung Kuo. Gaussian measures in Banach spaces. Springer-Verlag, Berlin, 1975. Lecture Notes in Mathematics, Vol. 463. [13] Peter Li and Shing-Tung Yau. On the parabolic kernel of the Schrödinger operator. Acta Math., 156(3-4):153–201, 1986. [14] Tai Melcher. Heat kernel analysis on semi-infinite Lie groups. J. Funct. Anal., 257(11):3552– 3592, 2009. [15] Feng-Yu Wang. Logarithmic Sobolev inequalities on noncompact Riemannian manifolds. Probab. Theory Related Fields, 109(3):417–424, 1997. [16] Feng-Yu Wang. Equivalence of dimension-free Harnack inequality and curvature condition. Integral Equations Operator Theory, 48(4):547–552, 2004. † Department of Mathematics, University of Connecticut, Storrs, CT 06269, U.S.A. E-mail address: [email protected]